feat(runtime): controllable thinking level for sessions #523

Background

The Maka runtime can already stream thinking content (ThinkingDeltaEvent) and marks reasoning-capable models (reasoning: true in model-catalog.ts), but it cannot control the thinking level. A reasoning model thinks at its default depth with no way to raise, lower, or turn it off. The CLI/TUI inherits this gap, which blocks dogfooding the TUI against reasoning models like GLM, Kimi, or DeepSeek.

Current State

• buildProviderOptions (packages/runtime/src/model-factory.ts) returns an empty options object per provider; it never sends reasoning_effort (OpenAI protocol) or thinking.budget_tokens (Anthropic protocol).
• There is no thinkingLevel field on session, settings, or the turn request.
• The TUI /model picker only switches model id; there is no thinking-variant or level switch.
• Pi (sibling @earendil-works/pi-coding-agent) already has a full thinking-level system: settings.defaultThinkingLevel, per-model thinkingLevelMap (off/minimal/low/medium/high/xhigh), and supportsReasoningEffort.

Goal

Let a user pick a thinking level per session from the TUI, mapped to each provider's reasoning-effort / thinking-budget parameter, with the chosen level shown in the statusline.

Proposed Plan

Phase 1 — Provider options

buildProviderOptions accepts a thinking level and emits the protocol-correct parameter:

• reasoning_effort for openai / openai-compatible / zai-coding-plan / ollama / ark-style providers
• thinking.budget_tokens for anthropic / claude-subscription / kimi-coding-plan / MiniMax

Phase 2 — Model metadata

Add a thinkingLevelMap to provider/model definitions (mirror Pi's ~/.pi/agent/models.json) so each level maps to the provider-specific value, e.g. { high: 'high', xhigh: 'max' }.

Phase 3 — Plumbing

Add thinkingLevel to session/settings and the turn request, thread it from the session driver through the backend to buildProviderOptions.

Phase 4 — TUI UI

/thinking <level> slash command (and/or a variant row in the model picker); show the current level in the statusline next to the model.

Phase 5 — Transcript

Render ThinkingDeltaEvent blocks (the event type already exists) with a hide toggle (Pi's hideThinkingBlock).

Non-Goals

• Auto-discovering whether a provider supports reasoning effort; rely on the declared thinkingLevelMap.
• Encoding thinking as a separate model id (e.g. glm-4.6-thinking); keep level as a session-level control.
• Per-message thinking-budget overrides in the first version; start with a per-session level.

Acceptance Criteria

• A /thinking command changes the session thinking level; subsequent turns send the mapped parameter.
• buildProviderOptions sends reasoning_effort for OpenAI-protocol providers and thinking.budget_tokens for Anthropic-protocol providers.
• A reasoning-capable model can be set to off (no thinking) and to higher levels, observable in the provider request.
• The statusline shows the current thinking level.
• Thinking blocks render in the transcript and can be collapsed/hidden.
• Existing provider option behavior (safety settings, store: false, etc.) is preserved.

References

• buildProviderOptions: packages/runtime/src/model-factory.ts
• ThinkingDeltaEvent / ThinkingCompleteEvent: packages/core/src/events.ts
• model reasoning capability: packages/core/src/model-catalog.ts
• provider catalog (zai-coding-plan, ollama, openai-compatible): packages/core/src/llm-connections.ts
• Pi reference config: ~/.pi/agent/models.json (thinkingLevelMap), ~/.pi/agent/settings.json (defaultThinkingLevel)